Re: [PATCH 0/4] x86/Hyper-V: Panic code path fixes

[Tianyu Lan <ltykernel@gmail.com>]

On 3/20/2020 12:07 AM, Michael Kelley wrote:
> From: Michael Kelley <mikelley@microsoft.com> Sent: Thursday, March 19, 2020 8:15 AM
>>
>> From: Tianyu Lan <Tianyu.Lan@microsoft.com> Sent: Thursday, March 19, 2020 7:08 AM
>>>>>
>>>>> This patchset fixes some issues in the Hyper-V panic code path.
>>>>> Patch 1 resolves issue that panic system still responses network
>>>>> packets.
>>>>> Patch 2-3 resolves crash enlightenment issues.
>>>>> Patch 4 is to set crash_kexec_post_notifiers to true for Hyper-V
>>>>> VM in order to report crash data or kmsg to host before running
>>>>> kdump kernel.
>>>>
>>>> I still see an issue that isn't addressed by these patches.  The VMbus
>>>> driver registers a "die notifier" and a "panic notifier".   But die() will
>>>> eventually call panic() if panic_on_oops is set (which I think it typically
>>>> is).  If the CRASH_NOTIFY_MSG option is *not* enabled, then
>>>> hyperv_report_panic() could get called by the die notifier, and then
>>>> again by the panic notifier.
>>>>
>>>> Do we even need the "die notifier"?  If it was removed, there would
>>>> not be any notification to Hyper-V via the die() path unless panic_on_oops
>>>> is set, which I think is actually the correct behavior.  I'm not
>>>> completely clear on what is supposed to happen in general to the
>>>> Linux kernel if panic_on_oops is not set. Does it try to continue to run?
>>>> If so, then we should not be notifying Hyper-V if panic_on_oops is not
>>>> set, and removing the die notifier is the right thing to do.
>>>>
>>>
>>> hyperv_report_panic() has re-enter check inside and so kernel only
>>> reports crash register data once during die().
>>
>> Ah, yes, you are right.
>>
>>>  From comment in the
>>> hyperv_report_panic(), register value reported in die chain is more
>>> exact than value in panic chain. The register value in die chain is
>>> passed by die() caller. Register value reported in panic chain
>>> is collected in the hyperv_panic_event().
>>>
>>> If panic_on_oops is not set, the task should be killed and kernel
>>> still runs. In this case, we may not trigger crash enlightenment.
>>
>> I'm not completely clear on your last statement.   It seems like there
>> is still a problem in that die() will call hyperv_report_panic() even if
>> panic_on_oops is not set.  We will have reported a panic to Hyper-V
>> even though the VM did not stop running.
>
Yes, the die callback is still necessary and we should skip report
if panic_on_oops isn't set.

> 
> There's one more issue to consider.  hv_kmsg_dump() skips calling
> hyperv_report_panic_msg() if sysctl_record_panic_msg has been cleared
> by a sysctl command.  (This sysctl option gives a customer the ability to
> increase privacy by not having the VM's dmesg contents sent to Hyper-V.)
> In this case, the earlier hyperv_report_panic() call should be used.  Otherwise,
> there would not be any notification to Hyper-V about the panic.
> 

Nice catch. I will fix this in the next version.

Thanks.